The model was trained in Darknet. It is based off of Tiny Yolov2, the only modification to the architecture was that the model only detects a single class instead of the default 80. It has been tested using Darkflow and is known to have good performance.
I first convert from Darknet to Tensorflow using Darkflow:
./flow --model cherry.cfg --load cherry_final.weights --savepb
This works and does not give me an error. It creates a .pb file and a .meta file. I next convert from Tensorflow to IR format using the model optimizer:
sudo python3 mo_tf.py --input_model cherry.pb --tensorflow_use_custom_operations_config cherry.json --input_shape "[1,1280,704,3]" --data_type FP16
This also works fine and does not return an error. It creates a .bin file and a .xml file. My json contents are as follows:
[
{
"id": "TFYOLO",
"match_kind": "general",
"custom_attributes": {
"classes": 1,
"coords": 4,
"num": 5,
"do_softmax": 1
}
}
]
Finally, I load my generated IR files and attempt to classify a sample image, using the following script:
from __future__ import division
import cv2 as cv
import numpy as np
import sys
np.set_printoptions(threshold=sys.maxsize)
# Load the model
net = cv.dnn.readNet('cherry.xml', 'cherry.bin')
# Specify target device
net.setPreferableTarget(cv.dnn.DNN_TARGET_MYRIAD)
# Read an image
frame = cv.imread('test.png')
img = (frame.copy()).astype(np.float32)
img = np.divide(img, 255)
img = img[:, :, ::-1]
input_image = img.astype(np.uint8)
# Prepare input blob and perform an inference
blob = cv.dnn.blobFromImage(input_image, size=(1280, 704), ddepth=cv.CV_8U)
net.setInput(blob)
out = net.forward()
detection = out.reshape(5,6,40,22)
for i in range(6):
for j in range(40):
for k in range(22):
confidence = float(detection[4,i,j,k])
xmin = int((detection[0,i,j,k] - detection[2,i,j,k]/2) * frame.shape[0])
ymin = int((detection[1,i,j,k] - detection[3,i,j,k]/2) * frame.shape[1])
xmax = int((detection[0,i,j,k] + detection[2,i,j,k]/2) * frame.shape[0])
ymax = int((detection[1,i,j,k] + detection[3,i,j,k]/2) * frame.shape[1])
if confidence > 0.8:
cv.rectangle(frame, (xmin, ymin), (xmax, ymax), color=(0, 255, 0))
Attempting to run this on an NCS2 results in the following error.
Traceback (most recent call last):
File "test_myriad_openvino.py", line 27, in <module>
out = net.forward()
cv2.error: OpenCV(4.0.1) /home/odroid/opencv-4.0.1/modules/dnn/src/op_inf_engine.cpp:555: error:
(-215:Assertion failed) Failed to initialize Inference Engine backend: [VPU] Internal error:
Output in 17-maxpool has incorrect width dimension. Expected: 21 or 21 Actual: 22 in function 'initPlugin'
If I run it on an NCS1, I don't get an error, but the detections it makes on the test image are not correct. The network should be able to identify people in the image. I have also tested with another image with only one person but I still get seemingly random results. Again this is a model that had no issues detecting people when tested in Darknet.
I tried running it on the CPU (cv.dnn.DNN_TARGET_CPU) but I get this error:
Traceback (most recent call last):
File "test_myriad_openvino.py", line 27, in <module>
out = net.forward()
cv2.error: OpenCV(4.0.1) /home/odroid/opencv-4.0.1/modules/dnn/src/op_inf_engine.cpp:555: error:
(-215:Assertion failed) in function 'initPlugin'
> Failed to initialize Inference Engine backend: Cannot find plugin to use :
>
Inferences are being run on Ubuntu 16.04, Openvino version is 5.455. Any assistance you can provide would be highly appreciated. Thanks.
Link Copied
